tensorflow 2.0

1-2 图片数据建模流程范例

tensorflow中准备图片数据的常用方案有两种，第一种是使用tf.keras中的ImageDataGenerator工具构建图片数据生成器。

第二种是使用tf.data.Dataset搭配tf.image中的一些图片处理方法构建数据管道。

第一种方法更为简单，其使用范例可以参考以下文章。

[《Keras图像数据预处理范例——Cifar2图片分类》](
第二种方法是TensorFlow的原生方法，更加灵活，使用得当的话也可以获得更好的性能。

我们此处介绍第二种方法。

import tensorflow as tf 
from tensorflow.keras import datasets,layers,models

BATCH_SIZE = 100

def load_image(img_path,size = (32,32)):
    label = tf.constant(1,tf.int8) if tf.strings.regex_full_match(img_path,".*automobile.*") \
            else tf.constant(0,tf.int8)
    img = tf.io.read_file(img_path)
    img = tf.image.decode_jpeg(img) #注意此处为jpeg格式
    img = tf.image.resize(img,size)/255.0
    return(img,label)

#使用并行化预处理num_parallel_calls 和预存数据prefetch来提升性能
ds_train = tf.data.Dataset.list_files("./data/cifar2/train/*/*.jpg") \
           .map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE) \
           .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
           .prefetch(tf.data.experimental.AUTOTUNE)  

ds_test = tf.data.Dataset.list_files("./data/cifar2/test/*/*.jpg") \
           .map(load_image, num_parallel_calls=tf.data.experimental.AUTOTUNE) \
           .batch(BATCH_SIZE) \
           .prefetch(tf.data.experimental.AUTOTUNE)  

%matplotlib inline
%config InlineBackend.figure_format = 'svg'

#查看部分样本
from matplotlib import pyplot as plt 

plt.figure(figsize=(8,8)) 
for i,(img,label) in enumerate(ds_train.unbatch().take(9)):
    ax=plt.subplot(3,3,i+1)
    ax.imshow(img.numpy())
    ax.set_title("label = %d"%label)
    ax.set_xticks([])
    ax.set_yticks([]) 
plt.show()

6-1,构建模型的3种方法

可以使用以下3种方式构建模型：使用Sequential按层顺序构建模型，使用函数式API构建任意结构模型，继承Model基类构建自定义模型。

对于顺序结构的模型，优先使用Sequential方法构建。

如果模型有多输入或者多输出，或者模型需要共享权重，或者模型具有残差连接等非顺序结构，推荐使用函数式API进行创建。

如果无特定必要，尽可能避免使用Model子类化的方式构建模型，这种方式提供了极大的灵活性，但也有更大的概率出错。

下面以IMDB电影评论的分类问题为例，演示3种创建模型的方法。

import numpy as np 
import pandas as pd 
import tensorflow as tf
from tqdm import tqdm 
from tensorflow.keras import *


train_token_path = "./data/imdb/train_token.csv"
test_token_path = "./data/imdb/test_token.csv"

MAX_WORDS = 10000  # We will only consider the top 10,000 words in the dataset
MAX_LEN = 200  # We will cut reviews after 200 words
BATCH_SIZE = 20 

# 构建管道
def parse_line(line):
    t = tf.strings.split(line,"\t")
    label = tf.reshape(tf.cast(tf.strings.to_number(t[0]),tf.int32),(-1,))
    features = tf.cast(tf.strings.to_number(tf.strings.split(t[1]," ")),tf.int32)
    return (features,label)

ds_train=  tf.data.TextLineDataset(filenames = [train_token_path]) \
   .map(parse_line,num_parallel_calls = tf.data.experimental.AUTOTUNE) \
   .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
   .prefetch(tf.data.experimental.AUTOTUNE)

ds_test=  tf.data.TextLineDataset(filenames = [test_token_path]) \
   .map(parse_line,num_parallel_calls = tf.data.experimental.AUTOTUNE) \
   .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
   .prefetch(tf.data.experimental.AUTOTUNE)

一，Sequential按层顺序创建模型

tf.keras.backend.clear_session()

model = models.Sequential()

model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
model.add(layers.MaxPool1D(2))
model.add(layers.Flatten())
model.add(layers.Dense(1,activation = "sigmoid"))

model.compile(optimizer='Nadam',
            loss='binary_crossentropy',
            metrics=['accuracy',"AUC"])

model.summary()

import datetime
baselogger = callbacks.BaseLogger(stateful_metrics=["AUC"])
logdir = "./data/keras_model/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
history = model.fit(ds_train,validation_data = ds_test,
        epochs = 6,callbacks=[baselogger,tensorboard_callback])

%matplotlib inline
%config InlineBackend.figure_format = 'svg'

import matplotlib.pyplot as plt

def plot_metric(history, metric):
    train_metrics = history.history[metric]
    val_metrics = history.history['val_'+metric]
    epochs = range(1, len(train_metrics) + 1)
    plt.plot(epochs, train_metrics, 'bo--')
    plt.plot(epochs, val_metrics, 'ro-')
    plt.title('Training and validation '+ metric)
    plt.xlabel("Epochs")
    plt.ylabel(metric)
    plt.legend(["train_"+metric, 'val_'+metric])
    plt.show()

plot_metric(history,"AUC")

二，函数式API创建任意结构模型

tf.keras.backend.clear_session()

inputs = layers.Input(shape=[MAX_LEN])
x  = layers.Embedding(MAX_WORDS,7)(inputs)

branch1 = layers.SeparableConv1D(64,3,activation="relu")(x)
branch1 = layers.MaxPool1D(3)(branch1)
branch1 = layers.SeparableConv1D(32,3,activation="relu")(branch1)
branch1 = layers.GlobalMaxPool1D()(branch1)

branch2 = layers.SeparableConv1D(64,5,activation="relu")(x)
branch2 = layers.MaxPool1D(5)(branch2)
branch2 = layers.SeparableConv1D(32,5,activation="relu")(branch2)
branch2 = layers.GlobalMaxPool1D()(branch2)

branch3 = layers.SeparableConv1D(64,7,activation="relu")(x)
branch3 = layers.MaxPool1D(7)(branch3)
branch3 = layers.SeparableConv1D(32,7,activation="relu")(branch3)
branch3 = layers.GlobalMaxPool1D()(branch3)

concat = layers.Concatenate()([branch1,branch2,branch3])
outputs = layers.Dense(1,activation = "sigmoid")(concat)

model = models.Model(inputs = inputs,outputs = outputs)

model.compile(optimizer='Nadam',
            loss='binary_crossentropy',
            metrics=['accuracy',"AUC"])

model.summary()

Model: "model"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 200)]        0                                            
__________________________________________________________________________________________________
embedding (Embedding)           (None, 200, 7)       70000       input_1[0][0]                    
__________________________________________________________________________________________________
separable_conv1d (SeparableConv (None, 198, 64)      533         embedding[0][0]                  
__________________________________________________________________________________________________
separable_conv1d_2 (SeparableCo (None, 196, 64)      547         embedding[0][0]                  
__________________________________________________________________________________________________
separable_conv1d_4 (SeparableCo (None, 194, 64)      561         embedding[0][0]                  
__________________________________________________________________________________________________
max_pooling1d (MaxPooling1D)    (None, 66, 64)       0           separable_conv1d[0][0]           
__________________________________________________________________________________________________
max_pooling1d_1 (MaxPooling1D)  (None, 39, 64)       0           separable_conv1d_2[0][0]         
__________________________________________________________________________________________________
max_pooling1d_2 (MaxPooling1D)  (None, 27, 64)       0           separable_conv1d_4[0][0]         
__________________________________________________________________________________________________
separable_conv1d_1 (SeparableCo (None, 64, 32)       2272        max_pooling1d[0][0]              
__________________________________________________________________________________________________
separable_conv1d_3 (SeparableCo (None, 35, 32)       2400        max_pooling1d_1[0][0]            
__________________________________________________________________________________________________
separable_conv1d_5 (SeparableCo (None, 21, 32)       2528        max_pooling1d_2[0][0]            
__________________________________________________________________________________________________
global_max_pooling1d (GlobalMax (None, 32)           0           separable_conv1d_1[0][0]         
__________________________________________________________________________________________________
global_max_pooling1d_1 (GlobalM (None, 32)           0           separable_conv1d_3[0][0]         
__________________________________________________________________________________________________
global_max_pooling1d_2 (GlobalM (None, 32)           0           separable_conv1d_5[0][0]         
__________________________________________________________________________________________________
concatenate (Concatenate)       (None, 96)           0           global_max_pooling1d[0][0]       
                                                                 global_max_pooling1d_1[0][0]     
                                                                 global_max_pooling1d_2[0][0]     
__________________________________________________________________________________________________
dense (Dense)                   (None, 1)            97          concatenate[0][0]                
==================================================================================================
Total params: 78,938
Trainable params: 78,938
Non-trainable params: 0
__________________________________________________________________________________________________

import datetime
logdir = "./data/keras_model/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
history = model.fit(ds_train,validation_data = ds_test,epochs = 6,callbacks=[tensorboard_callback])

Epoch 1/6
1000/1000 [==============================] - 32s 32ms/step - loss: 0.5527 - accuracy: 0.6758 - AUC: 0.7731 - val_loss: 0.3646 - val_accuracy: 0.8426 - val_AUC: 0.9192
Epoch 2/6
1000/1000 [==============================] - 24s 24ms/step - loss: 0.3024 - accuracy: 0.8737 - AUC: 0.9444 - val_loss: 0.3281 - val_accuracy: 0.8644 - val_AUC: 0.9350
Epoch 3/6
1000/1000 [==============================] - 24s 24ms/step - loss: 0.2158 - accuracy: 0.9159 - AUC: 0.9715 - val_loss: 0.3461 - val_accuracy: 0.8666 - val_AUC: 0.9363
Epoch 4/6
1000/1000 [==============================] - 24s 24ms/step - loss: 0.1492 - accuracy: 0.9464 - AUC: 0.9859 - val_loss: 0.4017 - val_accuracy: 0.8568 - val_AUC: 0.9311
Epoch 5/6
1000/1000 [==============================] - 24s 24ms/step - loss: 0.0944 - accuracy: 0.9696 - AUC: 0.9939 - val_loss: 0.4998 - val_accuracy: 0.8550 - val_AUC: 0.9233
Epoch 6/6
1000/1000 [==============================] - 26s 26ms/step - loss: 0.0526 - accuracy: 0.9865 - AUC: 0.9977 - val_loss: 0.6463 - val_accuracy: 0.8462 - val_AUC: 0.9138

plot_metric(history,"AUC")

三，Model子类化创建自定义模型

# 先自定义一个残差模块，为自定义Layer

class ResBlock(layers.Layer):
    def __init__(self, kernel_size, **kwargs):
        super(ResBlock, self).__init__(**kwargs)
        self.kernel_size = kernel_size
    
    def build(self,input_shape):
        self.conv1 = layers.Conv1D(filters=64,kernel_size=self.kernel_size,
                                   activation = "relu",padding="same")
        self.conv2 = layers.Conv1D(filters=32,kernel_size=self.kernel_size,
                                   activation = "relu",padding="same")
        self.conv3 = layers.Conv1D(filters=input_shape[-1],
                                   kernel_size=self.kernel_size,activation = "relu",padding="same")
        self.maxpool = layers.MaxPool1D(2)
        super(ResBlock,self).build(input_shape) # 相当于设置self.built = True
    
    def call(self, inputs):
        x = self.conv1(inputs)
        x = self.conv2(x)
        x = self.conv3(x)
        x = layers.Add()([inputs,x])
        x = self.maxpool(x)
        return x
    
    #如果要让自定义的Layer通过Functional API 组合成模型时可以序列化，需要自定义get_config方法。
    def get_config(self):  
        config = super(ResBlock, self).get_config()
        config.update({'kernel_size': self.kernel_size})
        return config

# 测试ResBlock
resblock = ResBlock(kernel_size = 3)
resblock.build(input_shape = (None,200,7))
resblock.compute_output_shape(input_shape=(None,200,7))

TensorShape([None, 100, 7])

# 自定义模型，实际上也可以使用Sequential或者FunctionalAPI

class ImdbModel(models.Model):
    def __init__(self):
        super(ImdbModel, self).__init__()
        
    def build(self,input_shape):
        self.embedding = layers.Embedding(MAX_WORDS,7)
        self.block1 = ResBlock(7)
        self.block2 = ResBlock(5)
        self.dense = layers.Dense(1,activation = "sigmoid")
        super(ImdbModel,self).build(input_shape)
    
    def call(self, x):
        x = self.embedding(x)
        x = self.block1(x)
        x = self.block2(x)
        x = layers.Flatten()(x)
        x = self.dense(x)
        return(x)

tf.keras.backend.clear_session()

model = ImdbModel()
model.build(input_shape =(None,200))
model.summary()

model.compile(optimizer='Nadam',
            loss='binary_crossentropy',
            metrics=['accuracy',"AUC"])

Model: "imdb_model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding (Embedding)        multiple                  70000     
_________________________________________________________________
res_block (ResBlock)         multiple                  19143     
_________________________________________________________________
res_block_1 (ResBlock)       multiple                  13703     
_________________________________________________________________
dense (Dense)                multiple                  351       
=================================================================
Total params: 103,197
Trainable params: 103,197
Non-trainable params: 0
_________________________________________________________________

import datetime

logdir = "./tflogs/keras_model/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
history = model.fit(ds_train,validation_data = ds_test,
                    epochs = 6,callbacks=[tensorboard_callback])

Epoch 1/6
1000/1000 [==============================] - 47s 47ms/step - loss: 0.5629 - accuracy: 0.6618 - AUC: 0.7548 - val_loss: 0.3422 - val_accuracy: 0.8510 - val_AUC: 0.9286
Epoch 2/6
1000/1000 [==============================] - 43s 43ms/step - loss: 0.2648 - accuracy: 0.8903 - AUC: 0.9576 - val_loss: 0.3276 - val_accuracy: 0.8650 - val_AUC: 0.9410
Epoch 3/6
1000/1000 [==============================] - 42s 42ms/step - loss: 0.1573 - accuracy: 0.9439 - AUC: 0.9846 - val_loss: 0.3861 - val_accuracy: 0.8682 - val_AUC: 0.9390
Epoch 4/6
1000/1000 [==============================] - 42s 42ms/step - loss: 0.0849 - accuracy: 0.9706 - AUC: 0.9950 - val_loss: 0.5324 - val_accuracy: 0.8616 - val_AUC: 0.9292
Epoch 5/6
1000/1000 [==============================] - 43s 43ms/step - loss: 0.0393 - accuracy: 0.9876 - AUC: 0.9986 - val_loss: 0.7693 - val_accuracy: 0.8566 - val_AUC: 0.9132
Epoch 6/6
1000/1000 [==============================] - 44s 44ms/step - loss: 0.0222 - accuracy: 0.9926 - AUC: 0.9994 - val_loss: 0.9328 - val_accuracy: 0.8584 - val_AUC: 0.9052

plot_metric(history,"AUC")

6-2,训练模型的3种方法

模型的训练主要有内置fit方法、内置tran_on_batch方法、自定义训练循环。

注：fit_generator方法在tf.keras中不推荐使用，其功能已经被fit包含。

import numpy as np 
import pandas as pd 
import tensorflow as tf
from tensorflow.keras import * 

#打印时间分割线
@tf.function
def printbar():
    today_ts = tf.timestamp()%(24*60*60)

    hour = tf.cast(today_ts//3600+8,tf.int32)%tf.constant(24)
    minite = tf.cast((today_ts%3600)//60,tf.int32)
    second = tf.cast(tf.floor(today_ts%60),tf.int32)
    
    def timeformat(m):
        if tf.strings.length(tf.strings.format("{}",m))==1:
            return(tf.strings.format("0{}",m))
        else:
            return(tf.strings.format("{}",m))
    
    timestring = tf.strings.join([timeformat(hour),timeformat(minite),
                timeformat(second)],separator = ":")
    tf.print("=========="*8+timestring)
    

MAX_LEN = 300
BATCH_SIZE = 32
(x_train,y_train),(x_test,y_test) = datasets.reuters.load_data()
x_train = preprocessing.sequence.pad_sequences(x_train,maxlen=MAX_LEN)
x_test = preprocessing.sequence.pad_sequences(x_test,maxlen=MAX_LEN)

MAX_WORDS = x_train.max()+1
CAT_NUM = y_train.max()+1

ds_train = tf.data.Dataset.from_tensor_slices((x_train,y_train)) \
          .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
          .prefetch(tf.data.experimental.AUTOTUNE).cache()
   
ds_test = tf.data.Dataset.from_tensor_slices((x_test,y_test)) \
          .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
          .prefetch(tf.data.experimental.AUTOTUNE).cache()

一，内置fit方法

该方法功能非常强大, 支持对numpy array, tf.data.Dataset以及 Python generator数据进行训练。

并且可以通过设置回调函数实现对训练过程的复杂控制逻辑。

tf.keras.backend.clear_session()
def create_model():
    
    model = models.Sequential()
    model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
    model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Flatten())
    model.add(layers.Dense(CAT_NUM,activation = "softmax"))
    return(model)

def compile_model(model):
    model.compile(optimizer=optimizers.Nadam(),
                loss=losses.SparseCategoricalCrossentropy(),
                metrics=[metrics.SparseCategoricalAccuracy(),metrics.SparseTopKCategoricalAccuracy(5)]) 
    return(model)
 
model = create_model()
model.summary()
model = compile_model(model)

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding (Embedding)        (None, 300, 7)            216874    
_________________________________________________________________
conv1d (Conv1D)              (None, 296, 64)           2304      
_________________________________________________________________
max_pooling1d (MaxPooling1D) (None, 148, 64)           0         
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 146, 32)           6176      
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 73, 32)            0         
_________________________________________________________________
flatten (Flatten)            (None, 2336)              0         
_________________________________________________________________
dense (Dense)                (None, 46)                107502    
=================================================================
Total params: 332,856
Trainable params: 332,856
Non-trainable params: 0
_________________________________________________________________

history = model.fit(ds_train,validation_data = ds_test,epochs = 10)

Train for 281 steps, validate for 71 steps
Epoch 1/10
281/281 [==============================] - 11s 37ms/step - loss: 2.0231 - sparse_categorical_accuracy: 0.4636 - sparse_top_k_categorical_accuracy: 0.7450 - val_loss: 1.7346 - val_sparse_categorical_accuracy: 0.5534 - val_sparse_top_k_categorical_accuracy: 0.7560
Epoch 2/10
281/281 [==============================] - 9s 31ms/step - loss: 1.5079 - sparse_categorical_accuracy: 0.6091 - sparse_top_k_categorical_accuracy: 0.7901 - val_loss: 1.5475 - val_sparse_categorical_accuracy: 0.6109 - val_sparse_top_k_categorical_accuracy: 0.7792
Epoch 3/10
281/281 [==============================] - 9s 33ms/step - loss: 1.2204 - sparse_categorical_accuracy: 0.6823 - sparse_top_k_categorical_accuracy: 0.8448 - val_loss: 1.5455 - val_sparse_categorical_accuracy: 0.6367 - val_sparse_top_k_categorical_accuracy: 0.8001
Epoch 4/10
281/281 [==============================] - 9s 33ms/step - loss: 0.9382 - sparse_categorical_accuracy: 0.7543 - sparse_top_k_categorical_accuracy: 0.9075 - val_loss: 1.6780 - val_sparse_categorical_accuracy: 0.6398 - val_sparse_top_k_categorical_accuracy: 0.8032
Epoch 5/10
281/281 [==============================] - 10s 34ms/step - loss: 0.6791 - sparse_categorical_accuracy: 0.8255 - sparse_top_k_categorical_accuracy: 0.9513 - val_loss: 1.9426 - val_sparse_categorical_accuracy: 0.6376 - val_sparse_top_k_categorical_accuracy: 0.7956
Epoch 6/10
281/281 [==============================] - 9s 33ms/step - loss: 0.5063 - sparse_categorical_accuracy: 0.8762 - sparse_top_k_categorical_accuracy: 0.9716 - val_loss: 2.2141 - val_sparse_categorical_accuracy: 0.6291 - val_sparse_top_k_categorical_accuracy: 0.7947
Epoch 7/10
281/281 [==============================] - 10s 37ms/step - loss: 0.4031 - sparse_categorical_accuracy: 0.9050 - sparse_top_k_categorical_accuracy: 0.9817 - val_loss: 2.4126 - val_sparse_categorical_accuracy: 0.6264 - val_sparse_top_k_categorical_accuracy: 0.7947
Epoch 8/10
281/281 [==============================] - 10s 35ms/step - loss: 0.3380 - sparse_categorical_accuracy: 0.9205 - sparse_top_k_categorical_accuracy: 0.9881 - val_loss: 2.5366 - val_sparse_categorical_accuracy: 0.6242 - val_sparse_top_k_categorical_accuracy: 0.7974
Epoch 9/10
281/281 [==============================] - 10s 36ms/step - loss: 0.2921 - sparse_categorical_accuracy: 0.9299 - sparse_top_k_categorical_accuracy: 0.9909 - val_loss: 2.6564 - val_sparse_categorical_accuracy: 0.6242 - val_sparse_top_k_categorical_accuracy: 0.7983
Epoch 10/10
281/281 [==============================] - 9s 30ms/step - loss: 0.2613 - sparse_categorical_accuracy: 0.9334 - sparse_top_k_categorical_accuracy: 0.9947 - val_loss: 2.7365 - val_sparse_categorical_accuracy: 0.6220 - val_sparse_top_k_categorical_accuracy: 0.8005

二，内置train_on_batch方法

该内置方法相比较fit方法更加灵活，可以不通过回调函数而直接在批次层次上更加精细地控制训练的过程。

tf.keras.backend.clear_session()

def create_model():
    model = models.Sequential()

    model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
    model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Flatten())
    model.add(layers.Dense(CAT_NUM,activation = "softmax"))
    return(model)

def compile_model(model):
    model.compile(optimizer=optimizers.Nadam(),
                loss=losses.SparseCategoricalCrossentropy(),
                metrics=[metrics.SparseCategoricalAccuracy(),metrics.SparseTopKCategoricalAccuracy(5)]) 
    return(model)
 
model = create_model()
model.summary()
model = compile_model(model)

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding (Embedding)        (None, 300, 7)            216874    
_________________________________________________________________
conv1d (Conv1D)              (None, 296, 64)           2304      
_________________________________________________________________
max_pooling1d (MaxPooling1D) (None, 148, 64)           0         
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 146, 32)           6176      
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 73, 32)            0         
_________________________________________________________________
flatten (Flatten)            (None, 2336)              0         
_________________________________________________________________
dense (Dense)                (None, 46)                107502    
=================================================================
Total params: 332,856
Trainable params: 332,856
Non-trainable params: 0
_________________________________________________________________

def train_model(model,ds_train,ds_valid,epoches):

    for epoch in tf.range(1,epoches+1):
        model.reset_metrics()
        
        # 在后期降低学习率
        if epoch == 5:
            model.optimizer.lr.assign(model.optimizer.lr/2.0)
            tf.print("Lowering optimizer Learning Rate...\n\n")
        
        for x, y in ds_train:
            train_result = model.train_on_batch(x, y)

        for x, y in ds_valid:
            valid_result = model.test_on_batch(x, y,reset_metrics=False)
            
        if epoch%1 ==0:
            printbar()
            tf.print("epoch = ",epoch)
            print("train:",dict(zip(model.metrics_names,train_result)))
            print("valid:",dict(zip(model.metrics_names,valid_result)))
            print("")

train_model(model,ds_train,ds_test,10)

================================================================================13:09:19
epoch =  1
train: {'loss': 0.82411176, 'sparse_categorical_accuracy': 0.77272725, 'sparse_top_k_categorical_accuracy': 0.8636364}
valid: {'loss': 1.9265995, 'sparse_categorical_accuracy': 0.5743544, 'sparse_top_k_categorical_accuracy': 0.75779164}

================================================================================13:09:27
epoch =  2
train: {'loss': 0.6006621, 'sparse_categorical_accuracy': 0.90909094, 'sparse_top_k_categorical_accuracy': 0.95454544}
valid: {'loss': 1.844159, 'sparse_categorical_accuracy': 0.6126447, 'sparse_top_k_categorical_accuracy': 0.7920748}

================================================================================13:09:35
epoch =  3
train: {'loss': 0.36935613, 'sparse_categorical_accuracy': 0.90909094, 'sparse_top_k_categorical_accuracy': 0.95454544}
valid: {'loss': 2.163433, 'sparse_categorical_accuracy': 0.63312554, 'sparse_top_k_categorical_accuracy': 0.8045414}

================================================================================13:09:42
epoch =  4
train: {'loss': 0.2304088, 'sparse_categorical_accuracy': 0.90909094, 'sparse_top_k_categorical_accuracy': 1.0}
valid: {'loss': 2.8911984, 'sparse_categorical_accuracy': 0.6344613, 'sparse_top_k_categorical_accuracy': 0.7978629}

Lowering optimizer Learning Rate...


================================================================================13:09:51
epoch =  5
train: {'loss': 0.111194365, 'sparse_categorical_accuracy': 0.95454544, 'sparse_top_k_categorical_accuracy': 1.0}
valid: {'loss': 3.6431572, 'sparse_categorical_accuracy': 0.6295637, 'sparse_top_k_categorical_accuracy': 0.7978629}

================================================================================13:09:59
epoch =  6
train: {'loss': 0.07741702, 'sparse_categorical_accuracy': 0.95454544, 'sparse_top_k_categorical_accuracy': 1.0}
valid: {'loss': 4.074161, 'sparse_categorical_accuracy': 0.6255565, 'sparse_top_k_categorical_accuracy': 0.794301}

================================================================================13:10:07
epoch =  7
train: {'loss': 0.056113098, 'sparse_categorical_accuracy': 1.0, 'sparse_top_k_categorical_accuracy': 1.0}
valid: {'loss': 4.4461513, 'sparse_categorical_accuracy': 0.6273375, 'sparse_top_k_categorical_accuracy': 0.79652715}

================================================================================13:10:17
epoch =  8
train: {'loss': 0.043448802, 'sparse_categorical_accuracy': 1.0, 'sparse_top_k_categorical_accuracy': 1.0}
valid: {'loss': 4.7687583, 'sparse_categorical_accuracy': 0.6224399, 'sparse_top_k_categorical_accuracy': 0.79741764}

================================================================================13:10:26
epoch =  9
train: {'loss': 0.035002146, 'sparse_categorical_accuracy': 1.0, 'sparse_top_k_categorical_accuracy': 1.0}
valid: {'loss': 5.130505, 'sparse_categorical_accuracy': 0.6175423, 'sparse_top_k_categorical_accuracy': 0.794301}

================================================================================13:10:34
epoch =  10
train: {'loss': 0.028303564, 'sparse_categorical_accuracy': 1.0, 'sparse_top_k_categorical_accuracy': 1.0}
valid: {'loss': 5.4559293, 'sparse_categorical_accuracy': 0.6148709, 'sparse_top_k_categorical_accuracy': 0.7947462}

三，自定义训练循环

自定义训练循环无需编译模型，直接利用优化器根据损失函数反向传播迭代参数，拥有最高的灵活性。

tf.keras.backend.clear_session()

def create_model():
    
    model = models.Sequential()

    model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
    model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Flatten())
    model.add(layers.Dense(CAT_NUM,activation = "softmax"))
    return(model)

model = create_model()
model.summary()

optimizer = optimizers.Nadam()
loss_func = losses.SparseCategoricalCrossentropy()

train_loss = metrics.Mean(name='train_loss')
train_metric = metrics.SparseCategoricalAccuracy(name='train_accuracy')

valid_loss = metrics.Mean(name='valid_loss')
valid_metric = metrics.SparseCategoricalAccuracy(name='valid_accuracy')

@tf.function
def train_step(model, features, labels):
    with tf.GradientTape() as tape:
        predictions = model(features,training = True)
        loss = loss_func(labels, predictions)
    gradients = tape.gradient(loss, model.trainable_variables)
    optimizer.apply_gradients(zip(gradients, model.trainable_variables))

    train_loss.update_state(loss)
    train_metric.update_state(labels, predictions)
    

@tf.function
def valid_step(model, features, labels):
    predictions = model(features)
    batch_loss = loss_func(labels, predictions)
    valid_loss.update_state(batch_loss)
    valid_metric.update_state(labels, predictions)
    

def train_model(model,ds_train,ds_valid,epochs):
    for epoch in tf.range(1,epochs+1):
        
        for features, labels in ds_train:
            train_step(model,features,labels)

        for features, labels in ds_valid:
            valid_step(model,features,labels)

        logs = 'Epoch={},Loss:{},Accuracy:{},Valid Loss:{},Valid Accuracy:{}'
        
        if epoch%1 ==0:
            printbar()
            tf.print(tf.strings.format(logs,
            (epoch,train_loss.result(),train_metric.result(),valid_loss.result(),valid_metric.result())))
            tf.print("")
            
        train_loss.reset_states()
        valid_loss.reset_states()
        train_metric.reset_states()
        valid_metric.reset_states()

train_model(model,ds_train,ds_test,10)

6-3,使用单GPU训练模型

深度学习的训练过程常常非常耗时，一个模型训练几个小时是家常便饭，训练几天也是常有的事情，有时候甚至要训练几十天。

训练过程的耗时主要来自于两个部分，一部分来自数据准备，另一部分来自参数迭代。

当数据准备过程还是模型训练时间的主要瓶颈时，我们可以使用更多进程来准备数据。

当参数迭代过程成为训练时间的主要瓶颈时，我们通常的方法是应用GPU或者Google的TPU来进行加速。

详见《用GPU加速Keras模型——Colab免费GPU使用攻略》

https://zhuanlan.zhihu.com/p/68509398

无论是内置fit方法，还是自定义训练循环，从CPU切换成单GPU训练模型都是非常方便的，无需更改任何代码。当存在可用的GPU时，如果不特意指定device，tensorflow会自动优先选择使用GPU来创建张量和执行张量计算。

但如果是在公司或者学校实验室的服务器环境，存在多个GPU和多个使用者时，为了不让单个同学的任务占用全部GPU资源导致其他同学无法使用（tensorflow默认获取全部GPU的全部内存资源权限，但实际上只使用一个GPU的部分资源），我们通常会在开头增加以下几行代码以控制每个任务使用的GPU编号和显存大小，以便其他同学也能够同时训练模型。

在Colab笔记本中：修改->笔记本设置->硬件加速器 中选择 GPU

注：以下代码只能在Colab 上才能正确执行。

可通过以下colab链接测试效果《tf_单GPU》：

https://colab.research.google.com/drive/1r5dLoeJq5z01sU72BX2M5UiNSkuxsEFe

%tensorflow_version 2.x
import tensorflow as tf
print(tf.__version__)

from tensorflow.keras import * 

#打印时间分割线
@tf.function
def printbar():
    today_ts = tf.timestamp()%(24*60*60)

    hour = tf.cast(today_ts//3600+8,tf.int32)%tf.constant(24)
    minite = tf.cast((today_ts%3600)//60,tf.int32)
    second = tf.cast(tf.floor(today_ts%60),tf.int32)
    
    def timeformat(m):
        if tf.strings.length(tf.strings.format("{}",m))==1:
            return(tf.strings.format("0{}",m))
        else:
            return(tf.strings.format("{}",m))
    
    timestring = tf.strings.join([timeformat(hour),timeformat(minite),
                timeformat(second)],separator = ":")
    tf.print("=========="*8+timestring)
    

一，GPU设置

gpus = tf.config.list_physical_devices("GPU")

if gpus:
    gpu0 = gpus[0] #如果有多个GPU，仅使用第0个GPU
    tf.config.experimental.set_memory_growth(gpu0, True) #设置GPU显存用量按需使用
    # 或者也可以设置GPU显存为固定使用量(例如：4G)
    #tf.config.experimental.set_virtual_device_configuration(gpu0,
    #    [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=4096)]) 
    tf.config.set_visible_devices([gpu0],"GPU") 

比较GPU和CPU的计算速度

printbar()
with tf.device("/gpu:0"):
    tf.random.set_seed(0)
    a = tf.random.uniform((10000,100),minval = 0,maxval = 3.0)
    b = tf.random.uniform((100,100000),minval = 0,maxval = 3.0)
    c = a@b
    tf.print(tf.reduce_sum(tf.reduce_sum(c,axis = 0),axis=0))
printbar()

================================================================================17:37:01
2.24953778e+11
================================================================================17:37:01

printbar()
with tf.device("/cpu:0"):
    tf.random.set_seed(0)
    a = tf.random.uniform((10000,100),minval = 0,maxval = 3.0)
    b = tf.random.uniform((100,100000),minval = 0,maxval = 3.0)
    c = a@b
    tf.print(tf.reduce_sum(tf.reduce_sum(c,axis = 0),axis=0))
printbar()

================================================================================17:37:34
2.24953795e+11
================================================================================17:37:40

二，准备数据

MAX_LEN = 300
BATCH_SIZE = 32
(x_train,y_train),(x_test,y_test) = datasets.reuters.load_data()
x_train = preprocessing.sequence.pad_sequences(x_train,maxlen=MAX_LEN)
x_test = preprocessing.sequence.pad_sequences(x_test,maxlen=MAX_LEN)

MAX_WORDS = x_train.max()+1
CAT_NUM = y_train.max()+1

ds_train = tf.data.Dataset.from_tensor_slices((x_train,y_train)) \
          .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
          .prefetch(tf.data.experimental.AUTOTUNE).cache()
   
ds_test = tf.data.Dataset.from_tensor_slices((x_test,y_test)) \
          .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
          .prefetch(tf.data.experimental.AUTOTUNE).cache()
          

三，定义模型

tf.keras.backend.clear_session()

def create_model():
    
    model = models.Sequential()

    model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
    model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Flatten())
    model.add(layers.Dense(CAT_NUM,activation = "softmax"))
    return(model)

model = create_model()
model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding (Embedding)        (None, 300, 7)            216874    
_________________________________________________________________
conv1d (Conv1D)              (None, 296, 64)           2304      
_________________________________________________________________
max_pooling1d (MaxPooling1D) (None, 148, 64)           0         
_________________________________________________________________
conv1d_1 (Conv1D)            (None, 146, 32)           6176      
_________________________________________________________________
max_pooling1d_1 (MaxPooling1 (None, 73, 32)            0         
_________________________________________________________________
flatten (Flatten)            (None, 2336)              0         
_________________________________________________________________
dense (Dense)                (None, 46)                107502    
=================================================================
Total params: 332,856
Trainable params: 332,856
Non-trainable params: 0
_________________________________________________________________

四，训练模型

optimizer = optimizers.Nadam()
loss_func = losses.SparseCategoricalCrossentropy()

train_loss = metrics.Mean(name='train_loss')
train_metric = metrics.SparseCategoricalAccuracy(name='train_accuracy')

valid_loss = metrics.Mean(name='valid_loss')
valid_metric = metrics.SparseCategoricalAccuracy(name='valid_accuracy')

@tf.function
def train_step(model, features, labels):
    with tf.GradientTape() as tape:
        predictions = model(features,training = True)
        loss = loss_func(labels, predictions)
    gradients = tape.gradient(loss, model.trainable_variables)
    optimizer.apply_gradients(zip(gradients, model.trainable_variables))

    train_loss.update_state(loss)
    train_metric.update_state(labels, predictions)
    
@tf.function
def valid_step(model, features, labels):
    predictions = model(features)
    batch_loss = loss_func(labels, predictions)
    valid_loss.update_state(batch_loss)
    valid_metric.update_state(labels, predictions)
    

def train_model(model,ds_train,ds_valid,epochs):
    for epoch in tf.range(1,epochs+1):
        
        for features, labels in ds_train:
            train_step(model,features,labels)

        for features, labels in ds_valid:
            valid_step(model,features,labels)

        logs = 'Epoch={},Loss:{},Accuracy:{},Valid Loss:{},Valid Accuracy:{}'
        
        if epoch%1 ==0:
            printbar()
            tf.print(tf.strings.format(logs,
            (epoch,train_loss.result(),train_metric.result(),valid_loss.result(),valid_metric.result())))
            tf.print("")
            
        train_loss.reset_states()
        valid_loss.reset_states()
        train_metric.reset_states()
        valid_metric.reset_states()

train_model(model,ds_train,ds_test,10)

6-4,使用多GPU训练模型

如果使用多GPU训练模型，推荐使用内置fit方法，较为方便，仅需添加2行代码。

在Colab笔记本中：修改->笔记本设置->硬件加速器 中选择 GPU

注：以下代码只能在Colab 上才能正确执行。

可通过以下colab链接测试效果《tf_多GPU》：

https://colab.research.google.com/drive/1j2kp_t0S_cofExSN7IyJ4QtMscbVlXU-

MirroredStrategy过程简介：

• 训练开始前，该策略在所有 N 个计算设备上均各复制一份完整的模型；
• 每次训练传入一个批次的数据时，将数据分成 N 份，分别传入 N 个计算设备（即数据并行）；
• N 个计算设备使用本地变量（镜像变量）分别计算自己所获得的部分数据的梯度；
• 使用分布式计算的 All-reduce 操作，在计算设备间高效交换梯度数据并进行求和，使得最终每个设备都有了所有设备的梯度之和；
• 使用梯度求和的结果更新本地变量（镜像变量）；
• 当所有设备均更新本地变量后，进行下一轮训练（即该并行策略是同步的）。

%tensorflow_version 2.x
import tensorflow as tf
print(tf.__version__)
from tensorflow.keras import * 

#此处在colab上使用1个GPU模拟出两个逻辑GPU进行多GPU训练
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
    # 设置两个逻辑GPU模拟多GPU训练
    try:
        tf.config.experimental.set_virtual_device_configuration(gpus[0],
            [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024),
             tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)])
        logical_gpus = tf.config.experimental.list_logical_devices('GPU')
        print(len(gpus), "Physical GPU,", len(logical_gpus), "Logical GPUs")
    except RuntimeError as e:
        print(e)

一，准备数据

MAX_LEN = 300
BATCH_SIZE = 32
(x_train,y_train),(x_test,y_test) = datasets.reuters.load_data()
x_train = preprocessing.sequence.pad_sequences(x_train,maxlen=MAX_LEN)
x_test = preprocessing.sequence.pad_sequences(x_test,maxlen=MAX_LEN)

MAX_WORDS = x_train.max()+1
CAT_NUM = y_train.max()+1

ds_train = tf.data.Dataset.from_tensor_slices((x_train,y_train)) \
          .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
          .prefetch(tf.data.experimental.AUTOTUNE).cache()
   
ds_test = tf.data.Dataset.from_tensor_slices((x_test,y_test)) \
          .shuffle(buffer_size = 1000).batch(BATCH_SIZE) \
          .prefetch(tf.data.experimental.AUTOTUNE).cache()

二，定义模型

tf.keras.backend.clear_session()
def create_model():
    
    model = models.Sequential()

    model.add(layers.Embedding(MAX_WORDS,7,input_length=MAX_LEN))
    model.add(layers.Conv1D(filters = 64,kernel_size = 5,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Conv1D(filters = 32,kernel_size = 3,activation = "relu"))
    model.add(layers.MaxPool1D(2))
    model.add(layers.Flatten())
    model.add(layers.Dense(CAT_NUM,activation = "softmax"))
    return(model)

def compile_model(model):
    model.compile(optimizer=optimizers.Nadam(),
                loss=losses.SparseCategoricalCrossentropy(from_logits=True),
                metrics=[metrics.SparseCategoricalAccuracy(),metrics.SparseTopKCategoricalAccuracy(5)]) 
    return(model)

三，训练模型

#增加以下两行代码
strategy = tf.distribute.MirroredStrategy()  
with strategy.scope(): 
    model = create_model()
    model.summary()
    model = compile_model(model)
    
history = model.fit(ds_train,validation_data = ds_test,epochs = 10)